BioPREP

This repository contains Dataset (BioPREP: Biomedical Predicate Relation Extraction with entity-filtering by PKDE4J) and code implementation for fine-tuning pretrained model, as well as inferencing unlabeled dataset using fine-tuned model.

Our open dataset, BioPREP, is based on SemMedDB (https://skr3.nlm.nih.gov/SemMed/index.html). We have extracted entities of sentences in SemMedDB using PKDE4J (link: http://informatics.yonsei.ac.kr/pkde4j/#), then replaced them with Entity Type. After preprocessing the dataset, we comprehensively evaluated performance of several architectures based on neural networks. Following the result of experiments, we found that BioBERT-based model outperformed other models for predicate classification.

For users who want to use BioBERT-based fine-tuned model, not going through the training process of pre-trained model, we uploaded fine-tuned model at Google Drive link. By executing test.py, you can just infer the predicate/frame type of unlabelled datasets.

Requirements

torch

tensorflow

keras

numpy

pandas

sklearn

How to Use

1. Git clone and Set the directory

git clone https://github.com/deeptextlab/BioPREP.git

2. Train models

You can either fine-tune BERT-based pretrained models(BERT-base, BioBERT and SciBERT), or train neural network-based models(CNN, Multichannel CNN, LSTM, BiLSTM and CNN + LSTM). We simply used the extension of pretrained BERT models for our research.

python3 train.py --model_type='BioBERT' \ 
		--label_type='Predicate' \
  		--seed=42 \
    		--epochs=20 \
      		--batch_size=16 \
        	--test_size=0.2 \
          	--max_len=512 \
                --lr=5e-5 \
              	--eps=1e-8 \
                --eval_interval=5 \
                --output_dir_path='/models' \
                --data_file_path='/BioPREP/train.csv'

Values for arguments above are default settings, so if you just run

python3 train.py

the training process will run with given default settings.

To train under diffenent conditions, change arguments as you want.

For argument 'model_type', you can select among 'BioBERT', 'SciBERT' and 'BERT-base'.

For argument 'label_type', you can choose between 'Predicate' and 'FrameNet'.

If you want to use your own dataset, please fit your columns into 'text', 'predicate_answer' and 'framenet_answer', as well as column names.

Note that you can also manage when the fine-tuned weights would be saved by using argument 'eval_interval'. For example, if you give value 5 for argument eval_interval, while setting epoch as 20, the training process would save the weight every 5 epoch: 5, 10, 15 and 20. You can also specify where to save your model by changing the value of argument 'output_dir_path'.

3. Inference with unlabelled dataset

With unlabelled dataset, you can infer answers when running the codes below.

python3 test.py --model_type='BioBERT' \ 
		--label_type='Predicate' \
      		--batch_size=16 \
          	--max_len=512 \
                --output_dir_path='/models' \
                --data_file_path='/BioPREP/train.csv'

If you want to use other settings, change the arguments.

Citation

If you want to use our dataset BioPREP, please cite as below.

@article{10.1016/j.jbi.2021.103888,
	author={Gibong Hong, Yuheun Kim, Yeonjung Choi, Min Song},
	(in press),
	title={{BioPREP}: Deep Learning-based Predicate Classification with SemMedDB},
	journal={Journal of Biomedical Informatics},
	year={2021},
	month={08},
	doi={10.1016/j.jbi.2021.103888},
	url={https://doi.org/10.1016/j.jbi.2021.103888}
	}

Contact Information

If you have any questions or issues, please contact Gibong Hong gibong.hong@yonsei.ac.kr or create a Github issue.